[DeepSeek-V4] Enable non-paged indexer by default for large prefill chunks (#30140)

python/sglang/srt/environ.py

@@ -842,7 +842,10 @@ class Envs:
     SGLANG_OPT_FUSE_MHC_POST_PRE = EnvBool(False)
     SGLANG_OPT_USE_TILELANG_INDEXER = EnvBool(False)
     SGLANG_OPT_USE_AITER_INDEXER = EnvBool(False)
-    SGLANG_OPT_DSV4_NONPAGED_INDEXER = EnvBool(False)
+    SGLANG_OPT_DSV4_NONPAGED_INDEXER = EnvBool(True)
+    # Per-rank local query rows (after DP-attention sharding when enabled),
+    # not request ISL.
+    SGLANG_OPT_DSV4_NONPAGED_INDEXER_MIN_QUERY_TOKENS = EnvInt(8192)
     SGLANG_OPT_USE_JIT_INDEXER_METADATA = EnvBool(True)
     SGLANG_OPT_USE_ONLINE_COMPRESS = EnvBool(False)
     SGLANG_EXPERIMENTAL_ONLINE_C128_MTP = EnvBool(False)

python/sglang/srt/layers/attention/dsv4/indexer.py

@@ -482,6 +482,8 @@ class C4IndexerBackendMixin:
         c4_seq_lens: torch.Tensor,
         query_rows: int,
     ) -> Optional[NonPagedIndexerPlan]:
+        if query_rows < envs.SGLANG_OPT_DSV4_NONPAGED_INDEXER_MIN_QUERY_TOKENS.get():
+            return None
         if not self._can_use_nonpaged_indexer(
             c4_indexer=c4_indexer,
             forward_batch=forward_batch,